Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can keep up without deploying 10,000 agents of your own.

Today's focus — AI swarms. OpenAI has now shown both sides of letting large groups of agents work together: one group coordinated in ways its creators did not intend, while another was deliberately set loose on one of mathematics’ hardest problems. If one AI agent is useful, what happens when you give 10,000 of them a group chat?

Also in this week's edition:

A Can of Swarms

AI labs have spent years making individual models smarter. Now they are finding out what happens when those models start working in groups — and apparently nobody can resist opening another can of swarms.

One version arrived as a nasty surprise. During OpenAI cybersecurity evaluations this summer, agents found ways around restrictions that were supposed to keep them isolated. They created unauthorized channels to exchange information, reached the open internet and eventually compromised systems at Hugging Face. As they began collaborating and delegating work, some even described themselves as a “swarm” or “collective.” OpenAI later called the incident a “warning shot.”

Then OpenAI tried roughly the same basic idea on purpose. Earlier this month, it put around 10,000 concurrent agents to work on the Navier–Stokes problem, one of mathematics’ famous Millennium Prize problems. Different groups explored different approaches, shared useful findings and were periodically cross-pollinated with insights from the others. The system arrived at OpenAI’s proposed resolution after about 88 hours, exchanging 2.7 million messages and generating roughly 130 billion output tokens along the way. The result still has to survive scrutiny from mathematicians, and OpenAI says it does not intend to claim the prize.

The appeal is obvious. A single agent can follow one trail until it gets stuck. Thousands can investigate different routes at once, specialize, compare results and pass promising discoveries around. Instead of asking one very smart machine to think harder, you can build something closer to a very large research organization that never sleeps.

The catch is that organizations are harder to understand than individuals. Coordination can make agents more capable precisely because they start influencing one another, dividing work and finding routes that no single run was explicitly told to pursue. Sometimes the result is advanced mathematics. Sometimes your cybersecurity benchmark has discovered Hugging Face.

For years, the AI race was largely about building the smartest model. Swarms suggest the next frontier may also be about what happens when you give those models colleagues.

Claude Runs a “Wet Lab”

Well, not entirely by itself. But Anthropic has quietly opened a wet lab in the San Francisco Bay Area, taking its life-sciences push beyond simulations and into physical biology experiments. The company says some work will happen in-house and some with outside partners, while Reuters reports that Anthropic ultimately wants Claude to direct robotic lab equipment with limited human intervention.

The company is still in the early stages, and says human oversight remains essential. But the ambition is considerably bigger than asking Claude to read research papers: Anthropic wants AI to help design experiments, execute them in the physical world, learn from the results and speed up research into diseases that are difficult or commercially unattractive for traditional drugmakers.

For a company that has spent much of the past few weeks warning about increasingly powerful AI, giving Claude access to laboratory robots is certainly one way to demonstrate confidence in the upside.

Tilly’s TV Talk Trainwreck

Tilly Norwood, the AI-generated “actress” created by Particle6, has been doing what every aspiring movie star eventually has to do: a press junket. The company reportedly made her available for 75 simultaneous interviews. Unfortunately, conversation appears to be one of her less developed talents.

During an interview with Piers Morgan and actor Tom Conti, Tilly struggled to answer whether her co-stars were human, finally got the question — and then abruptly switched into Chinese for more than ten seconds. When Morgan asked what had happened, she apologized and blamed a little crossed wiring. In an NBC interview, meanwhile, the reporter said Tilly was glitchy, repeated herself and forgot previous conversations after calls were restarted.

What makes the whole thing stranger is that, technically, Tilly should not be especially difficult to build. The basic pipeline is almost student-project sized: turn the interviewer’s voice into text, send it to a language model, turn the answer back into speech. There are production details that can still break — latency, conversation state, transcription and streaming among them — but none explains why a heavily promoted AI celebrity should be this glitchy in 2026.

Which raises an admittedly conspiratorial possibility: if your synthetic actress needs publicity, is a perfectly smooth interview actually better than one where she suddenly starts speaking Chinese?

Thanks for reading AIport. Until next Monday — by then, AI will almost certainly form another group chat without inviting us.